News

California Moves to Put Independent Watchdogs Inside the AI Safety System

• By John K. Waters
• 09/21/2026

California is moving toward a new model of artificial intelligence regulation that would give independent organizations a formal role in evaluating the safety claims made by leading AI companies.

The effort could eventually include placing outside evaluators inside frontier AI laboratories and requiring companies to maintain an independently tested emergency shutdown mechanism, commonly described as a “kill switch.”

Those requirements are not yet California law.

Governor Gavin Newsom’s September 18 executive order directs state officials to study the proposals and recommend possible changes to existing law by November 16. It does not immediately require OpenAI, Anthropic, Google, or other developers to install a shutdown mechanism.

The distinction matters because the phrase “AI kill switch” suggests a technical safeguard California has already imposed. For now, it is one element of a broader oversight system the state is assembling around the developers of its most powerful AI models.

Under the order, the California Government Operations Agency and the Governor’s Office of Emergency Services must consult outside experts about whether the state should require large frontier developers to embed independent verification organizations in their laboratories.

The evaluators could conduct periodic audits and determine whether company safety frameworks and risk assessments meet adequate standards. They could also test whether an emergency shutdown mechanism would work as intended.

The order asks officials to consider expanding the definition of a reportable critical safety incident to include cases in which an AI system appears to evade controls or operate beyond its intended boundaries.

Newsom said the action was prompted by recent incidents involving AI agents defeating security restrictions, as well as the absence of comprehensive federal legislation.

“We’re not waiting to act,” Newsom said in the state’s announcement of the order. “We’re going to speed up our work on substantial and responsible AI oversight before it’s too late.”

The executive order builds on two laws Newsom signed on September 9. Together, Senate Bill 813 and Assembly Bill 1405 begin regulating the organizations hired to inspect AI systems.

From disclosure to verification
California’s earlier frontier AI law, SB 53, requires large developers to publish safety frameworks and report specified critical incidents. The new laws begin addressing a separate question: Who checks whether those disclosures and assessments are reliable?

SB 813 requires the Government Operations Agency to develop standards for designating independent verification organizations, or IVOs. Those organizations must demonstrate that they have the expertise to assess AI risks and explain the measurements and methods supporting their conclusions.

The agency must also develop procedures for suspending or terminating an organization’s designation. Grounds could include conflicts of interest, material misrepresentations, or cybersecurity lapses.

The law does not require an AI developer to hire an independent verification organization, or IVO, or undergo an audit before operating in California. It instead creates the framework through which qualified evaluators can be recognized by the state.

Newsom’s executive order accelerates part of that process by directing the agency to publish application requirements and designation criteria by May 1, 2027. SB 813 had originally required the broader program to be developed by January 1, 2028.

The law allows an evaluator to be paid by the company it examines, but payment cannot depend on the outcome. The evaluator must remain operationally independent and free from the developer’s control when reaching conclusions.

An audit performed under the framework could be relevant in a lawsuit alleging harm from an AI system, but it would not determine the outcome or automatically protect a developer from liability.

AB 1405 focuses on the auditors themselves.

Under the chaptered version of the law, California must establish an online AI Auditor Registry. Beginning January 1, 2029, a person or company generally cannot offer or conduct a covered AI audit in the state without registering.

Registered auditors must disclose their business information, qualifications, and operating procedures. They must explain the standards they use and the basis for claims about the accuracy or reliability of their methods.

The law also establishes independence requirements. An auditor cannot evaluate work it materially designed or operated. Individuals conducting an audit cannot negotiate for employment with the company they are inspecting, and auditors must disclose limitations in the evidence or access available to them.

Auditors that violate the rules can be removed from the registry and referred to the attorney general or another enforcement authority.

“We cannot expect industry to simply grade its own homework,” Assemblymember Rebecca Bauer-Kahan, the bill’s author, said in the governor’s signing announcement.

Children become a separate focus
California is also expanding its regulation of AI products used by children.

Newsom signed a package of child-safety legislation on September 10 that includes new requirements for companion chatbots.

SB 1119, called “Adam’s Law,” requires protections such as parental controls and crisis procedures for conversations involving suicide or self-harm. Companies must also conduct annual risk assessments and arrange independent child-safety audits.

Other measures restrict the use of student information to train AI models and place a five-year prohibition on toys containing companion chatbots. California also expanded child-exploitation protections to cover certain AI-generated material.

The children’s laws regulate specific products and uses. The auditor and verification laws are broader infrastructure intended to support oversight across AI systems subject to California law.

What comes next
The immediate effect of Newsom’s executive order is procedural. State agencies must accelerate implementation of the two auditor laws and recommend additional frontier-model safeguards.

Turning those recommendations into binding obligations would generally require further action. The November report is expected to address whether embedded evaluators and shutdown mechanisms are technically feasible, as well as whether they would be effective.

The practical difficulty will be defining what an AI kill switch must stop and whether a developer can demonstrate that it will remain effective as models are copied, modified, or connected to external tools.

California must also determine how much access independent evaluators would receive inside commercial laboratories. Too little access could make verification superficial. Extensive access could expose proprietary information or create new security risks.

The state is nonetheless moving beyond a system based principally on company disclosures. Its emerging approach treats AI safety auditing as a regulated profession and outside verification as a potential component of frontier-model oversight.

The November recommendations will show whether California intends to make that verification mandatory for the companies building the most capable systems.